Contents lists available at ScienceDirect Energy Conversion and Management journal homepage: www.elsevier.com/locate/enconman An effective secondary decomposition approach for wind power forecasting using extreme learning machine trained by crisscross optimization Hao Yin a , Zhen Dong a , Yunlong Chen a , Jiafei Ge a , Loi Lei Lai a , Alfredo Vaccaro b , Anbo Meng a, ⁎ a School of Automation, Guangdong University of Technology, Guangzhou 510006, Guangdong, China b Engineering Department, University of Sannio, Benevento 82100, Italy ARTICLE INFO Keywords: Wind power forecasting Secondary hybrid decomposition Empirical mode decomposition Wavelet packet decomposition Extreme learning machine Crisscross optimization algorithm ABSTRACT Large-scale integration of wind energy into electric grid is restricted by its inherent intermittence and volatility. So the increased utilization of wind power necessitates its accurate prediction. The contribution of this study is to develop a new hybrid forecasting model for the short-term wind power prediction by using a secondary hybrid decomposition approach. In the data pre-processing phase, the empirical mode decomposition is used to de- compose the original time series into several intrinsic mode functions (IMFs). A unique feature is that the generated IMF1 continues to be decomposed into appropriate and detailed components by applying wavelet packet decomposition. In the training phase, all the transformed sub-series are forecasted with extreme learning machine trained by our recently developed crisscross optimization algorithm (CSO). The final predicted values are obtained from aggregation. The results show that: (a) The performance of empirical mode decomposition can be significantly improved with its IMF1 decomposed by wavelet packet decomposition. (b) The CSO algorithm has satisfactory performance in addressing the premature convergence problem when applied to optimize ex- treme learning machine. (c) The proposed approach has great advantage over other previous hybrid models in terms of prediction accuracy. 1. Introduction Different from the conventional power that is dispatchable and easy to predict, the wind power has its inherent nature of volatility and in- termittence. Due to the uncertainty of wind power generation, the large-scale integration of wind turbines into power system is restricted [1]. Therefore, accurate wind power prediction is of great significance for power system operation in terms of unit commitment, energy market efficiency as well as lowering the cost by reducing the power reserves [2]. In the past few decades, several wind power forecasting approaches have been proposed, which usually fall into physical, empirical and artificial intelligence (AI) methods. The physical method predicts wind power by utilizing the numerical weather prediction (NWP) data into the manufacturer power curves [3]. However, the physical method is very complicated and is not reliable for short-term prediction. So it is usually used as an input for empirical models [4]. The empirical methods aim to describe the relation between historical time series of wind power at the location of interest by generally recursive techniques [5]. Most of the empirical models, such as autoregressive moving average model (ARMA) [6], autoregressive integrated moving average model (ARIMA) [7], assume that the wind speed data is normally dis- tributed. However, it is a well known characteristic of general wind speed series that its variation at a given site can be modeled using the Weibull distribution, which is not a normally distributed function and as a result, a transformation of the original wind speed data is required making the time series unstable and difficult to predict [8]. In addition, the volatility of wind power time series requires more complex function for capturing the stochastic relations, but these models are based on the assumption that a linear correlation structure exists among time series values [9]. In recent years, many machine learning forecasting techniques have been developed to address the nonlinear time series-based wind power forecasting problem. Among them, the artificial neural network (ANN) has become a popular method for wind energy forecasting due to its ability to capture the nonlinear relationship among the historical data. The applications of different ANNs in the wind power prediction field can be found in [10–13]. Compared with traditional algorithms such as the BP (back-propagation), the extreme learning machine (ELM) is a powerful algorithm with faster learning speed and better performance [14]. ELM tries to get the smallest training error and norm of weights. More examples of applying ELM to wind power prediction can be found http://dx.doi.org/10.1016/j.enconman.2017.08.014 Received 10 March 2017; Received in revised form 1 July 2017; Accepted 4 August 2017 ⁎ Corresponding author. E-mail address: menganbo@vip.sina.com (A. Meng). Energy Conversion and Management 150 (2017) 108–121 0196-8904/ © 2017 Elsevier Ltd. All rights reserved. MARK